Retailers face forecasting challenges at different levels of granularity, from individual store orders to supplier funding across stores and category performance by store. A new arXiv paper tackles this hierarchy by proposing a forecasting approach that blends world models, tree-based methods, and foundation models.
The work is motivated by the fact that decisions at each grain require different types of predictions and constraints. The authors test their method on four public data sets, providing evidence for its applicability across varied retail scenarios.
Because the abstract is truncated, specific performance numbers and comparisons are not available. The paper's contribution lies in framing retail forecasting as a multi-grain problem and demonstrating a combined modeling strategy on public benchmarks.